RyanStudio/Discord-Dialogues-Filtered

Discord Dialogues Filtered

I filtered the mookiezi/Discord-Dialogues dataset to obtain only high-quality english conversation examples for fine-tuning or other analytical tasks.

The total data set went from 7,303,464 rows to 2,208 rows after strict filtering to remove the following:

• Low conversation turns or short conversations
• Non-English conversations
• Duplicate conversations
• Spammy/Repetitive conversations
• Filler words

Dataset Statistics

Metric	Total	Avg / row	Median / row
Tokens (cl100k_base)	333,300	151.0	144
Words	260,590	118.0	115
Characters	1,364,613	618.0	598
Turns	15,932	7.22	6

Turn stats: min 4, max 24. Interquartile range of chars/row: 545–672 (P25–P75); tokens/row: 134–160 (P25–P75). Median single-turn length: 80 chars.

Role balance: 7,465 user turns (46.9%) vs 8,467 assistant turns (53.1%) across 15,932 total turns.

Rules

Phase 1 — Load, Parse & Size Filter

The raw dataset, containing 7,303,464 samples, is loaded and each conversation's text field is parsed into structured messages. Message counts (turns) and token counts (tokens, via cl100k_base) are recomputed so later stages can reason about conversation size.

Conversations are kept only if they meet both of the following:

• At least 4 turns
• Between 128 and 2,048 tokens

Rows removed: 7,301,008 (7,303,464 → 2,456) — the vast majority of the raw dataset was filtered out here for being too short, too long, or unparseable.

---

Phase 2 — Language Filter

Each conversation's joined text is passed through a fast language detector. Only conversations detected as English with sufficient confidence are retained.

A row is kept only when both hold:

• Detected language is en
• Detector confidence ≥ 0.80

Rows removed: 211 (2,456 → 2,245)

---

Phase 3 — Near-Duplicate Removal (MinHash LSH)

Exact and near-duplicate conversations are identified using character-5-gram MinHash signatures (256 permutations, bucketed into 32 bands of 8 rows). A row is dropped if it shares a band bucket with an already-kept row whose estimated Jaccard similarity is ≥ 0.70. Exact duplicates are also caught via SHA-256 digest in the same pass. When duplicates exist, the longest copy is kept.

Rows removed: 3 (2,245 → 2,242)

---

Phase 4 — Semantic Near-Duplicate Removal

Conversations that are paraphrases or rewordings of each other — but not necessarily string-similar — are removed using all-MiniLM-L6-v2 embeddings and a 2048-bit random-hyperplane LSH (128 bands of 16 bits). A row is dropped when its exact cosine similarity to an already-kept row is ≥ 0.80, gated by a Hamming pre-filter to cheaply prune unrelated candidates. As in Phase 3, the longest copy among duplicates is kept.

Rows removed: 0 (2,242 → 2,242)

---

Phase 5 — Filler & Spam Filter

Conversations dominated by low-information or spammy content are removed, while noisy turns inside otherwise-good conversations are preserved. Each row is scored against the rules below, and any rule firing triggers a drop. Dropped rows are written to dropped.jsonl along with the triggering reasons.

Rule	Threshold	What it catches
too_short	total words < 20	conversations too thin to be useful
high_filler	filler words / total words ≥ 0.50	small-talk / interjection chatter
low_unique	distinct words / total words < 0.25	repetitive spam
repetition	most-frequent word / total words ≥ 0.55	single-word spam
punct_spam	non-alphanumeric chars / total chars ≥ 0.20	..... spam
char_run	chars inside 5+ char runs / total chars ≥ 0.20	PPPP... / EEEE... spam
dup_messages	repeated message contents / total messages ≥ 0.50	repeated turns
single_word	one-word turns / total turns ≥ 0.60	boop / blep / kk spam
short_turns	median turn length < 6 chars	ultra-short ok / hm chatter

Rows removed: 34 (2,242 → 2,208)

Phase 5 drop-reason breakdown

Of the 34 dropped rows (some triggered multiple rules):

5  char_run
5  low_unique + repetition
4  low_unique + dup_messages
4  too_short
4  too_short + char_run
3  low_unique
3  dup_messages
3  punct_spam
1  too_short + char_run + single_word
1  single_word + short_turns
1  dup_messages + single_word + short_turns

---

Cumulative Summary

Phase	Input	Output	Removed
1 — Size filter	7,303,464	2,456	7,301,008
2 — Language	2,456	2,245	211
3 — Near-duplicate	2,245	2,242	3
4 — Semantic dedupe	2,242	2,242	0
5 — Filler / spam	2,242	2,208	34

Total removed: 7,301,256 rows (~99.97% of the raw dataset). Final clean dataset: 2,208 high-quality conversations.